LSTM层输入维度不兼容报错：预期ndim=3，实际得到ndim=2

问题背景 #

拥有含69个特征、225700行的数据集，运行LSTM自编码器代码时出现以下报错：

ValueError: Input 0 of layer "lstm_3" is incompatible with the layer: expected ndim=3, found ndim=2. Full shape received: (None, 68)

用户代码 #

import pandas as pd
from keras.layers import Input, LSTM, RepeatVector
from keras.models import Model

DF = pd.read_csv(r"C:\Users\44759\All_Autoencoder_Data.csv")
DF1 = DF.drop('Labels', axis=1) # 移除标签特征

# 定义输入序列形状
input_seq_shape = (DF1.shape[1], 1)

print(input_seq_shape) # 输出: (68, 1)

# 定义LSTM自编码器模型
inputs = Input(shape=input_seq_shape)
encoded = LSTM(68, activation='relu')(inputs)
encoded = LSTM(32, activation='relu')(encoded)
encoded = LSTM(5, activation='relu')(encoded)

decoded = RepeatVector(DF1.shape[1])(encoded)
decoded = LSTM(5, activation='relu', return_sequences=True)(decoded)
decoded = LSTM(32, activation='relu', return_sequences=True)(decoded)
decoded = LSTM(68, activation='relu', return_sequences=True)(decoded)
decoded = LSTM(1, activation='sigmoid', return_sequences=True)(decoded)

报错详情 #

ValueError                                Traceback (most recent call last)
<ipython-input-6-cf499e4225da> in <module>
      2 inputs = Input(shape=input_seq_shape)
      3 encoded = LSTM(68, activation='relu')(inputs)
----> 4 encoded = LSTM(32, activation='relu')(encoded)
      5 encoded = LSTM(5, activation='relu')(encoded)
      6 

~\anaconda3\lib\site-packages\keras\layers\rnn\base_rnn.py in __call__(self, inputs, initial_state, constants, **kwargs)
    554 
    555         if initial_state is None and constants is None:
---> 556             return super().__call__(inputs, **kwargs)
    557 
    558         # If any of `initial_state` or `constants` are specified and are Keras

~\anaconda3\lib\site-packages\keras\utils\traceback_utils.py in error_handler(*args, **kwargs)
     68             # To get the full stack trace, call:
     69             # `tf.debugging.disable_traceback_filtering()`
---> 70             raise e.with_traceback(filtered_tb) from None
     71         finally:
     72             del filtered_tb

~\anaconda3\lib\site-packages\keras\engine\input_spec.py in assert_input_compatibility(input_spec, inputs, layer_name)
    230             ndim = shape.rank
    231             if ndim != spec.ndim:
---> 232                 raise ValueError(
    233                     f'Input {input_index} of layer "{layer_name}" '
    234                     "is incompatible with the layer: "

ValueError: Input 0 of layer "lstm_3" is incompatible with the layer: expected ndim=3, found ndim=2. Full shape received: (None, 68)

问题原因 #

LSTM层默认参数return_sequences=False，此时输出是2D张量，形状为(batch_size, units)。而后续的LSTM层要求输入必须是3D张量，形状为(batch_size, timesteps, features)。

代码中第一层LSTM输出2D张量(None,68)，直接传给第二层LSTM时，就会触发维度不匹配的报错。

解决方法 #

在编码部分的前两层LSTM中添加return_sequences=True，让它们返回完整的序列（3D张量），只有最后一层编码LSTM保持默认（return_sequences=False），输出2D张量给RepeatVector层使用。

修改后的编码部分代码：

# 定义LSTM自编码器模型
inputs = Input(shape=input_seq_shape)
# 前两层LSTM添加return_sequences=True
encoded = LSTM(68, activation='relu', return_sequences=True)(inputs)
encoded = LSTM(32, activation='relu', return_sequences=True)(encoded)
# 最后一层编码LSTM不需要返回序列
encoded = LSTM(5, activation='relu')(encoded)

decoded = RepeatVector(DF1.shape[1])(encoded)
decoded = LSTM(5, activation='relu', return_sequences=True)(decoded)
decoded = LSTM(32, activation='relu', return_sequences=True)(decoded)
decoded = LSTM(68, activation='relu', return_sequences=True)(decoded)
decoded = LSTM(1, activation='sigmoid', return_sequences=True)(decoded)

额外注意事项 #

确保输入数据已经调整为符合LSTM要求的3D形状：(样本数, 时间步长, 特征数)。代码中input_seq_shape=(68,1)，对应每个样本是68个时间步、每个时间步1个特征，需要提前将DF1转换为该形状：

# 转换输入数据形状
X = DF1.values.reshape(-1, DF1.shape[1], 1)

内容的提问来源于stack exchange，提问作者Frenzy